Change default config.

We found that networks without normalization seem to have better alpha mask quality. Also, it is claimed in some literatures that networks w/o normalization preserve style information and networks with normalizaiton preserve content information.
Saeed-1986 · Sep 7, 2018 · f26d198 · f26d198
1 parent f67d114
commit f26d198
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/colab_demo/faceswap-GAN_colab_demo.ipynb b/colab_demo/faceswap-GAN_colab_demo.ipynb
@@ -491,7 +491,7 @@
     "arch_config = {}\n",
     "arch_config['IMAGE_SHAPE'] = (RESOLUTION, RESOLUTION, 3)\n",
     "arch_config['use_self_attn'] = True\n",
-    "arch_config['norm'] = \"instancenorm\" # instancenorm, batchnorm, layernorm, groupnorm, none\n",
+    "arch_config['norm'] = \"hybrid\" # instancenorm, batchnorm, layernorm, groupnorm, none\n",
     "arch_config['model_capacity'] = \"lite\" # standard, lite\n",
     "\n",
     "# Loss function weights configuration\n",